Update README.md

README.md

@@ -11,5 +11,127 @@ tags:
 - bitcoin
 pretty_name: Multi‑Timeframe Market Regimes (HMM‑6) (BTCUSD)
 ---
-# Dataset Summary
-Labeled crypto market regimes derived from multi‑timeframe OHLCV features and technical indicators (5m, 15m, 1h). Labels come from a 6‑state Hidden Markov Model (HMM). Useful for regime detection and sequence modeling baselines (LSTM/Transformers). No future leakage: indicators computed only up to timestamp and labels correspond to the regime at that timestamp.
+
+# Multi‑Timeframe Market Regimes (HMM‑6)
+
+## Dataset Summary
+
+Labeled crypto market regimes derived from multi‑timeframe OHLCV features and technical indicators (5m, 15m, 1h). Labels come from a 6‑state Hidden Markov Model (HMM). Useful for regime detection and sequence modeling baselines (LSTM/Transformers). **No future leakage**: indicators computed only up to `timestamp` and labels correspond to the regime at that timestamp.
+
+## Source & Licensing
+
+* **Raw data**: describe the exchange(s), symbols (e.g., `BTCUSDT`), and the exact ToS allowing redistribution. Replace this section with precise citations/links.
+* **License**: change the license above if the data source requires a specific license. If redistribution isn’t allowed, provide a script to regenerate labels from the original source instead of shipping raw candles.
+
+## Files
+
+```
+.
+├── README.md
+└── data/
+    ├── train.csv         # chronological split (older → newer)
+    ├── validation.csv    # next chunk after train
+    └── test.csv          # most recent chunk
+```
+
+Optionally provide Parquet equivalents:
+
+```
+└── data/
+    ├── train.parquet
+    ├── validation.parquet
+    └── test.parquet
+```
+
+## Columns
+
+* `timestamp` \[UTC, ISO8601 or UNIX ms]
+* **Multi‑TF OHLCV (5m, 15m, 1h)**: `open_*`, `high_*`, `low_*`, `close_*`, `volume_*`
+* **Indicators (15m)**: `log_ret_1_15m`, `ema_slope_21_15m`, `price_vs_ema55_15m`, `macd_hist_15m`, `adx_15m`, `atr_norm_15m`, `bb_width_15m`, `realized_vol_20_15m`, `rsi_15m`, `roc_15m`, `stoch_k_15m`, `wick_ratio_15m`
+* **Indicators (5m)**: `log_ret_1_5m`, `ema_slope_21_5m`, `price_vs_ema55_5m`, `macd_hist_5m`, `adx_5m`, `atr_norm_5m`, `bb_width_5m`, `realized_vol_20_5m`, `rsi_5m`, `roc_5m`, `stoch_k_5m`, `wick_ratio_5m`
+* **Indicators (1h)**: `log_ret_1_1h`, `ema_slope_21_1h`, `price_vs_ema55_1h`, `macd_hist_1h`, `adx_1h`, `atr_norm_1h`, `bb_width_1h`, `realized_vol_20_1h`, `rsi_1h`, `roc_1h`, `stoch_k_1h`, `wick_ratio_1h`
+* **Targets**: `state` (0‑5), `regime` (string label, optional), `post_prob_*` (optional soft posteriors per state)
+
+## Label Semantics
+
+```
+0: Choppy High‑Vol
+1: Range
+2: Squeeze
+3: Strong Trend
+4: Volatility Spike
+5: Weak Trend
+```
+
+* `state` is the integer id.
+* `regime` is the human‑readable label (optional redundant column).
+* If available, include HMM posterior probabilities per state: `post_prob_0` … `post_prob_5`.
+
+## Splitting Protocol (leak‑free)
+
+* Sort by `timestamp` ascending.
+* Choose three **contiguous** time blocks: `train` (oldest), `validation` (middle), `test` (most recent).
+* Do **not** shuffle.
+* If training sequence models, consumers should build sliding windows of length `time_steps` (e.g., 64) **within** each split only.
+
+## How to Load
+
+Python (CSV):
+
+```python
+from datasets import load_dataset
+# If you keep the README YAML above, this works without specifying data_files
+ds = load_dataset("<user_or_org>/<repo>")
+train = ds["train"]
+valid = ds["validation"]
+# Or explicitly with Parquet
+ds = load_dataset("<user_or_org>/<repo>", data_files={
+    "train": "data/train.parquet",
+    "validation": "data/validation.parquet",
+    "test": "data/test.parquet",
+})
+```
+
+PyTorch example (windowing):
+
+```python
+import torch
+import numpy as np
+FEATURES = [
+    "log_ret_1_15m","ema_slope_21_15m","price_vs_ema55_15m","macd_hist_15m","adx_15m","atr_norm_15m","bb_width_15m","realized_vol_20_15m","rsi_15m","roc_15m","stoch_k_15m","wick_ratio_15m",
+    "log_ret_1_5m","ema_slope_21_5m","price_vs_ema55_5m","macd_hist_5m","adx_5m","atr_norm_5m","bb_width_5m","realized_vol_20_5m","rsi_5m","roc_5m","stoch_k_5m","wick_ratio_5m",
+    "log_ret_1_1h","ema_slope_21_1h","price_vs_ema55_1h","macd_hist_1h","adx_1h","atr_norm_1h","bb_width_1h","realized_vol_20_1h","rsi_1h","roc_1h","stoch_k_1h","wick_ratio_1h"
+]
+TARGET = "state"
+WINDOW = 64
+
+def make_windows(table):
+    X = np.stack([table[f] for f in FEATURES], axis=1)
+    y = np.array(table[TARGET])
+    Xw, yw = [], []
+    for i in range(len(X) - WINDOW + 1):
+        Xw.append(X[i:i+WINDOW])
+        yw.append(y[i+WINDOW-1])  # label at window end
+    return np.stack(Xw), np.array(yw)
+```
+
+## Dataset Card Checklist
+
+* [ ] Precise data source + collection dates + timezone
+* [ ] License & ToS compliance
+* [ ] Exact indicator definitions (lookback windows, smoothing, normalization)
+* [ ] HMM configuration (n\_states=6, emissions, training window, seed)
+* [ ] Train/val/test date ranges
+* [ ] Class distribution per split
+* [ ] Known caveats & failure modes
+* [ ] Version tag (e.g., v1.0.0) and changelog
+
+## Known Limitations
+
+* Technical‑indicator engineered features; not raw trades.
+* Single asset (if BTCUSDT only) may not generalize to alts/FX.
+* Regimes are **model‑dependent**; HMM re‑fits may shift boundaries.
+
+## Changelog
+
+* v1.0.0 — Initial release.